Researchers at the University of Waterloo in Canada have made a breakthrough in artificial intelligence (AI). New machine learning algorithms, along with compact computer chips small enough to fit on a mobile phone, have allowed for an AI to run independent of the internet.
Typical neural networks (known as “tethered” neural networks) require an internet connection to run an AI program. This breakthrough, however, allows devices to operate an AI independent of the internet.
Having a stand-alone AI, as opposed to a tethered AI, lowers data transmission cost, increases privacy and security, and creates opportunities for the employment of AI in areas without internet connectivity.
Deep-learning AI emulates the human brain by processing data through layers of artificial neurons. Ordinarily, this requires massive computational power and data resources. The researchers at the University of Waterloo, however, were able to increase the efficiency of the AI by placing it in a virtual environment and slowly reducing the computational power and data available to it, thereby training it to use the remaining resources more efficiently.
“There are [a] huge number of situations where you simply cannot depend on the cloud for doing deep-learning AI computing, due to constraints on internet bandwidth, privacy and security issues, the lack of internet accessibility, as well as mission-critical tasks where reliable, real-time AI is mandatory,” said Alex Wong, associate professor in the Department of Systems Design Engineering at the University of Waterloo and co-creator of this technology.
“All of these pretty much boils down to the lack of available resources, and so that inspired us to investigate what happens when you build a virtual environment with very little resources and see how the AI adapts to it,” Wang said. “We see this form of evolved adaptation in nature all the time, where one’s physical characteristics adapt over time to the environment one is in, and so we wanted to see if that will happen to deep-learning AI as well.”
Wong and his team applied the Darwinian concept of “survival of the fittest” to train their AI to operate at higher levels of efficiency by incrementally decreasing the resources available to it. Basically, they taught it to conserve and budget by increasing resource scarcity.
Going forward, the researchers want to perfect the Darwinian machine-learning process they developed and then bring the resultant AI program to market.
“Our next step is to see if we can push this idea even further in the academic realm, and investigate how we can modify the virtual environment to get the desired evolution characteristics we are looking for,” said Wong. “We are also commercializing our efficient yet powerful deep-learning AI solutions through a startup called DarwinAI, as University of Waterloo is a hotbed for entrepreneurial innovation.”
The researchers believe their technology could open up the AI world.
“We feel that this can be game-changing for those who wish to take the big ideas in deep-learning AI and get it into a form that can be used by the masses,” Wong said.
Essentially we hope that this will help further democratize AI for anyone, anywhere, and anytime.
Mohammad Javad Shafiee, research assistant professor in the Department of Systems Design Engineering at the university, is the technology’s co-creator and DarwinAI’s co-founder.